Model Card for Susant-Achary/smolLM2-recipes

This is a 360-million-parameter causal language model, fine-tuned with LoRA adapters on a cleaned Indian recipe dataset. It generates step-by-step cooking instructions in English.

Model Details

Model Description

A compact transformer-decoder causal language model (360M parameters) originally from HuggingFaceTB/SmolLM2-360M, enhanced with low-rank adapter finetuning (LoRA) on ~22K Indian recipes (EmTpro01/indian-recipe-cleaned). Ideal for generating detailed, structured recipe instructions.

• Developed by: Susant Achary
• Model type: Causal Language Model (decoder-only transformer)
• Language(s): English
• License: Apache-2.0
• Finetuned from: HuggingFaceTB/SmolLM2-360M

Model Sources

• Hugging Face Hub: https://huggingface.co/Susant-Achary/smolLM2-recipes
• Dataset: EmTpro01/indian-recipe-cleaned on Hugging Face Datasets

Uses

Direct Use

• Generate detailed, step-by-step Indian recipe instructions given a recipe title and ingredient list.
• Integrate into cooking chatbots or recipe suggestion systems.

Downstream Use

• Fine-tune further for multilingual or specialized dietary versions.
• Plug into kitchen assistant apps or voice-controlled cooking aids.

Out-of-Scope Use

• Nutritional analysis or precise caloric computation (model does not predict nutritional values).
• Medical or allergy advice—recommend consulting a qualified professional.

Bias, Risks, and Limitations

The model reflects the distribution of the training data (predominantly North and South Indian recipes). It may reproduce cultural biases (e.g., under-representation of certain regional cuisines) and occasional inconsistencies in ingredient quantities or cooking times.

Recommendations

• Validate generated recipes before use.
• Provide user feedback loops to catch and correct errors.
• Augment with nutritional or safety checks from external sources.

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

tokenizer = AutoTokenizer.from_pretrained("Susant-Achary/smolLM2-recipes")
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained("Susant-Achary/smolLM2-recipes", device_map="auto")

generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
prompt = (
    "### Recipe: Masala Dosa\n"
    "### Ingredients:\n"
    "rice, urad dal, fenugreek seeds, potato, onion, mustard seeds, curry leaves\n\n"
    "### Instructions:\n"
)
print(generator(prompt, max_length=200, do_sample=True, top_p=0.9))